JP2020004391A - Method for obtaining at least one significant feature in a series of components of the same type and method for the classification of a component of such a series - Google Patents

Abstract

A method for acquiring at least one important feature in a series of components of the same type based on a data set acquired by non-destructive inspection.
Inspection of a classified random sample by a non-destructive inspection method by obtaining a 3D data set in each case for each component, and a plurality of non-defective and non-defective components in the random sample And extracting a defect-free component region from all of the components in the random sample; and at least one characteristic characteristic of the component type and the manufacturing process of the component. Determining, using a statistical method from the field of data analysis or a predetermined set of features, a feature that indicates a significant difference between good and bad products over time in manufacturing; Defining a feature and its attributes as a trained classifier.
[Selection diagram] Fig. 1

Description

  The present invention provides a method for acquiring at least one important feature in a series of components of the same type based on a data set acquired by non-destructive inspection, and a method for acquiring at least one important feature in a series of components of the same type. A method for classifying two components on the basis of a data set obtained by non-destructive inspection, evaluating the manufacturing process and in particular detecting trends and influencing the process in a proactive manner The purpose is to be able to give.

  Methods used to date to evaluate and control manufacturing processes with various process parameters using computed tomography data rely exclusively on information obtained from defects or discontinuities. These are observations, each compiled from a separate evaluation, and therefore do not create a relationship between a series of manufactured components (populations). In addition, it should be noted that most of the data obtained is not incorporated into the evaluation at all, especially in the case of computed tomography (CT). These data are all areas in the component without discontinuities and defects. In the field of CT, this is typically 90-95% of the data generated during inspection of each component being examined. In the case of components that are not defective, these known methods do not yield any further information other than the information that the component is OK.

  X-ray CT enables a three-dimensional display of components examined using this method, in particular of internal structures. The internal structure of the component may vary as a result of many effects on the process, which results may be reflected in structural changes. These changes differ in nature depending on the manufacturing process and on the material or combination of materials used. Depending on the nature of the changes, these are widely acceptable, but can be visualized with X-ray CT.

  Accordingly, it is an object of the present invention to provide a method that allows the system to be trained and subsequently inspected for such components, wherein discontinuities that lead to failure of the manufactured components during the inspection. And to provide a method for detecting variations and above all trends in the manufacturing process at an early stage.

  This object is achieved according to the invention by a method having the features of claims 1 and 5. Advantageous embodiments are specified in the dependent claims.

According to these, the object is a method for obtaining at least one important feature in a series of components of the same type on the basis of a data set obtained by non-destructive inspection, comprising: Law,
Examining a sorted random sample of a plurality of components whose production order is known by a non-destructive inspection method by obtaining a 3D dataset in each case for each component;
Separating the plurality of components in the random sample into non-defective products and non-defective products;
Extracting a defect-free component region from all of the components in the random sample;
At least one characteristic characteristic of the type of component and the manufacturing process of the component, the significant difference between good and bad in that characteristic over time in the manufacture of the plurality of components; At least one of the features from the field of data analysis or from a predetermined set of features, by neural network, machine learning approaches such as multi-instance learning, or by genetic programming or conventional statistical methods. Determining using a statistical method;
Defining at least one feature and its characteristics as a trained classifier.

  This method represents a first step designed as a learning phase. The starting point is a conditioned random sample, a series of components whose order of manufacture is known. The essential subdivision required for conditioning in the learning phase consists in separating the random sample into good and bad (OK / NOK). In the first step of the method, for example, a machine learning method is used. Machine learning methods automatically identify features that are characteristic of the process and type of component, and that exhibit significant differences between good and bad products. Alternatively, the features can be selected from a predetermined set.

  In an advantageous development of the invention, in addition to the last step according to the invention (the step of defining at least one characteristic as a trained classifier), at least one process parameter is integrated into the production of the component. Determining whether there is a correlation between the process parameter and at least one feature, and if so, at least one feature and its attributes are defined as a trained classifier and have no correlation If the above correlation determination is made for another process parameter until there is a correlation, then at least one feature is defined as a trained classifier and no correlation is found for any of the process parameters Means that each step of claim 1 is performed on another feature, and then the steps in claim It is provided that is repeated the characteristics of said another. If the above process parameters are incorporated into the learning phase, then in principle, the possibility of creating a direct correlation between the features identified in the volume data set and the process parameters can arise. Therefore, control of the process parameters based on the X-ray image information and automatic optimization are possible.

  In a further advantageous development of the invention, it is provided that a plurality of features are determined according to step d of claim 1 and that a combination of at least two features produces a correlation with at least one process parameter. In principle, a change in one process parameter is not only accompanied by a change in one feature, but also more complex, so that a change in one process parameter affects several features. In addition, the combination of features may increase the importance of correlation.

  For example, the following are used as process parameters: manufacturing process pressure, temperature, electrical and mechanical stress or flow, ambient temperature or air humidity. It is ensured that these process parameters are assigned (individually) one to one during the production of the component.

  In a further advantageous development of the invention, it is provided that the method is performed only on spatial regions of the component that can be predetermined. This reduces the size of the volume data considered for later steps, thereby speeding up the process.

According to claim 6, the object is a method for classifying one component in a series of components of the same type based on a data set obtained by a non-destructive inspection method, comprising: Law,
i) The field of data analysis using trend analysis for the presence of the development of characteristic features of defective products, using neural networks, machine learning approaches such as multi-instance learning, or genetic programming or conventional statistical methods. Using the trained classifier obtained with reference to any one of claims 1 to 5 to perform using a statistical method from the or from a predetermined set of features; ,
ii) in signaling that no correction is necessary in the manufacturing process of the component when no such development is detected, or in communicating a warning and / or obtaining a trained classifier Performing automatic corrective actions in the component manufacturing process by modifying the process parameters via an interface with the manufacturing machine, if the process parameters were used.

  The second phase (after the learning phase) is to determine if there is any process degradation during the manufacturing process. Such deterioration of the process may lead to deterioration of the component parts, and in the worst case, may lead to the failure of the component parts. In some cases, however, the deterioration may be caused by changing related process parameters. Can be countered, and in some cases this can be done automatically. This characteristic feature determined by the learning phase is used. The quality features of the various components, their attributes and possibly also the regions within the components are also identified in the learning phase, and in the second phase the application of the data learned in the first phase is performed. (Thus, this second phase is also called the apply phase). In the case of a series of examinations, all components arranged for examination are examined by computed tomography. The results of the feature determination are evaluated by trend analysis, along with any process parameters that may be available. This trend analysis is based on, for example, machine learning or uses a statistical method. If a tendency to develop towards the characteristic features of the defective product (NOK) is detected, then a regulatory intervention in the process can be initiated. This intervention depends on the results of the trend analysis and can be performed automatically if the process parameters are part of the training data set and can be detected continuously in the application phase. Therefore, according to the present invention, it is possible to intervene in the process, thereby preventing the production of defective products. This allows prediction of the process behavior with a corresponding proactive reaction.

  In an advantageous development of the invention, it is provided that at least one defect-free component region is extracted from the examined components, and that the method for classification is performed only on the defect-free component region. Provided. In such an approach, regions within the component are segmented to increase the importance of the feature. This step makes it possible to reduce the size of the volume data, which is considered in particular in later steps, and significantly reduces the time required for the evaluation.

  In a further advantageous development of the invention, it is provided that the non-destructive inspection method performed is a computed tomography, in particular an X-ray tomography. A particular advantage of computed tomography is the high density of information obtained, which increases the likelihood of finding important features. Instead of X-ray tomography, a method using ultrasonic waves or millimeter waves, terahertz spectroscopy or thermography can also be used.

  Further details and advantages of the present invention are explained in more detail below with reference to exemplary embodiments represented in the drawings.

FIG. 4 is a flowchart of a learning phase according to the present invention. FIG. 7 is a flowchart of an application phase according to the present invention after executing a learning phase.

  FIG. 1 shows the principle flow of the method in the learning phase according to the invention. This learning phase will be described in more detail below.

  In order to acquire features, first, a classified random sample is required (step number 1). That is, in the case of a simple two-class problem, for example, it is divided into a “good” class and a “bad” class. However, it is also possible to solve the problem of working with more than two classifications. A random sample requires a sufficient number of elements. The number of elements depends on the complexity of the classification task, and hundreds of elements can produce good results. However, in each case, a million (partial) data set may be required. Each element needs to be uniquely assigned to a classification so that the classification system can be trained.

  In step 2, a component area having no defect is extracted. In each case, an area within the component may be excluded based on a quality standard defined for the component to be examined. For example, if the porosity is critical to the quality of the component, the component area containing the porosity is not used to acquire features. Other additional quality criteria for pore size can be, for example, geometric deviation from nominal size or drawing specification (CAD), component cracking, component material separation, shrinkage holes, contamination.

  Extraction of a component region without defects can be performed manually, with the operator visually determining whether the extracted region is defect-free. Since there is an automatic image processing method that can detect defects for all the above-mentioned quality criteria, automatic extraction and evaluation are also possible. Thus, the method of extracting a component region without defects can be designed to extract the region step-by-step from the volume data set of the components of the region and evaluate the region using an image processing method. If a defect is detected, the area is discarded and not used for training. Regions that have not been rejected through image processing are incorporated into training or application.

  In step 3, an inquiry is made as to whether a predetermined feature should be selected.

  If the answer to this is negative, according to step number 4, the neural network or machine learning approach determines the inspection decisions made on the component based on the classified random samples and the structural characteristics inside the component. Offers the possibility to automatically determine features that indicate a correlation between In such an approach, models such as multi-instance learning, among others, are used to maximize learning outcomes.

  In contrast, if the answer to the query at step 3 is affirmative, then according to step number 5, a predetermined feature is provided. To this end, conventional statistical methods from the field of data analysis can be used to identify the features that generate the above-mentioned correlations. A prerequisite is that these features be pre-selected. This can be done manually by skilled image processing professionals or through automated processes used for image processing (using genetic programming).

  Then, according to step number 6 (following step number 4 or 5), the selection of important features and / or component regions is performed. This can be done, for example, via a neural network or data analysis. In addition to the type of feature, local density distribution, fiber orientation or local wall thickness variations, the location in the component where the characteristic feature is determined are also important to show the above correlation. possible. Limiting one or more features to a particular area / region of a component may contribute to the importance of the features.

  At step number 7, it is queried whether additional environmental data should be sent to the training of the classification system.

  If the answer to this is affirmative, in step number 8 the correlation between the process parameters and the characteristics of the features is determined. This is done using neural networks or data analysis as well. If the features that ultimately enable the evaluation of the process quality have been identified, the task set at this step is to find a match between specific process parameters and the characteristics of the feature. For example, if the pressure is high, the analysis may determine that a particular feature takes a particular value or forms a pattern. Thereby, a correlation between the characteristic feature and one or more process parameters is generated. If this feature is to be detected during operation, for example, a pressure adjustment may be introduced as a measure.

  Following step number 8 or if the answer to the inquiry of step number 7 is negative, a classifier trained at step number 9 is established. Once the learning process is complete, there will be a trained network, for example when a neural network is used (filter setting and weighting for each neuron in the network). The network and its parameters represent a classifier. In the case of predetermined features, representing the knowledge of the trained classifier is a characteristic feature that is important for each classification (optionally, the location in the component where this feature is evaluated). Related).

  FIG. 2 shows the principle flow of the method in the application phase according to the invention. This application phase is described in more detail below.

  If the important features, their specific attributes, and optionally their local relevance, are selected through the method of the learning phase described above (step no. 9), they are in normal operation (sequence of tests). ). Here, either all components (in-line operation) or associated random specimens (eg, every 10 components) are examined using X-ray CT according to step number 10.

  A component region having no defect is extracted according to step number 11 in the same manner as in the training.

  Optionally, according to step number 12, the selection of the relevant component area is also made based on the result of the learning phase.

  Using the trained classifier previously established in the learning phase according to step 9, in step 13 the features and classification are then calculated. For this purpose, previously identified features are evaluated in the possibly relevant component areas. The sum of the individual results (each feature within the various evaluated component areas) is then evaluated using a classification method (eg, using a neural network) and according to step number 14, a negative tendency is determined. A determination is made as to whether it has been detected.

  If no negative trend is detected at step number 14, this determination follows step number 15 and indicates to the operator that no correction is needed.

  If a negative tendency is detected at step number 14, an inquiry is made as to whether the process parameters have been learned according to step number 16.

  In the simplest case, that is, when the answer to this inquiry is negative, according to step number 17, a warning is given about the process change. This means that the process is likely to generate defective products.

  If the learned process parameters are present, the answer to the query at step number 16 is affirmative, which may further indicate the correlation between the features that can predict that a discontinuity will occur and the potential causes. . Such environmental data can be key influencing factors such as manufacturing machine parameters (pressure, temperature, stress, flow), but secondary effects such as, for example, ambient temperature, air humidity and operators. It can also be a factor. Critical to finding features during the learning phase is that the importance over time of the manufacture of the component is detected. Only through the temporal element can conclusions be made about the detection of process behaviors and trends, which are ultimately a prerequisite for a proactive approach to avoiding defects.

  If additional environmental parameters are involved, according to step number 18, the method is used to automatically execute the process and the identified process parameters (eg, pressure) are changed via the classification system. . Thereby, the negative tendency of the manufacturing process is canceled, and the formation of defective products is prevented in advance.

  Thus, the present invention can be summarized as follows:

  The purpose of the method according to the invention is not to detect the exceeding of the limit for unacceptable discontinuities, but to use the variation between components within an acceptable range to control the process. is there. This method requires a series of non-destructive inspections, and in the example embodiment, computer tomography is used for this. In this way, for each inspected component, a 3D data set that reproduces the internal structure of the component is generated. Through this series of inspections, a change in the manufacturing process can be detected in principle. Thereby, a tendency having a predictive property is visualized. Thus, it is possible to anticipate process changes, thereby counteracting process deterioration through intervention.

  An essential part of this method is that the information is obtained exclusively from discontinuities or defect-free component areas. There are various features that can be correlated with component quality and process quality, depending on the nature of the components and variations in the process.

  A decisive difference between the present invention and the prior art method is that the method of the present invention mainly relies only on information obtained from defect-free areas of the component. However, it is also possible to use information from additional defect prone areas in order to increase the importance of the characteristic features. In a method of extracting information from a component defect, meaningful information cannot be obtained from a component having no defect in terms of process evaluation and process influence. However, in the case of the prior art method, there is a monitoring gap, since trends can be visible even in the case of components without defects. Therefore, it is not possible to manufacture with zero defects because defects are always required to obtain information. Therefore, such a method is left behind exclusively. In contrast, the method according to the invention described herein does not require fault-prone components to take action in the application phase, so it is necessary to take the initiative and deal with it. Can be.

Claims (8)

A method for obtaining at least one important feature in a series of components of the same type based on a dataset obtained by non-destructive inspection, said non-destructive inspection comprising:
a) examining a sorted random sample of a plurality of components whose production order is known by a non-destructive inspection method by obtaining a 3D dataset in each case for each component;
b) dividing the plurality of components in the random sample into non-defective products and non-defective products;
c) extracting a defect-free component area from all of the component parts in the random sample;
d) at least one characteristic characteristic of the type of component and the manufacturing process of the component, wherein the quality of the plurality of components is significant between good and bad over time over the course of manufacturing. At least one feature that exhibits a significant difference from the field of data analysis or from a predetermined set of features, by neural network, machine learning approaches such as multi-instance learning, or by genetic programming or conventional statistical methods. Determining using the statistical method of
e) defining the at least one feature and its attributes as a trained classifier.   In addition to step e, it is determined whether at least one process parameter has been incorporated into the manufacture of the component and there is a correlation between the process parameter and the at least one feature and its attributes. In some cases, the at least one feature and its attributes are defined as a trained classifier, and if there is no correlation, the aforementioned correlation determination is made for another process parameter until there is a correlation, and thereafter 2. If the at least one feature is defined as a trained classifier and no correlation is found for any of the process parameters, each step of claim 1 is performed on another feature, 2. The method of claim 1, wherein steps in this claim are repeated for the other feature.   3. A method according to claim 1 or claim 2, wherein a plurality of features are determined according to step d of claim 1, and a combination of at least two features produces a correlation with at least one process parameter.   4. The method according to claim 2, wherein pressure, temperature, stress or flow of the manufacturing process, ambient temperature or air humidity are used as process parameters.   The method according to any of the preceding claims, wherein the method is performed only on a spatial region of the component that can be predetermined. A method for classifying one component in a series of components of the same type based on a data set obtained by a non-destructive inspection method, wherein the non-destructive inspection method includes:
i) The field of data analysis using a neural network, a machine learning approach such as multi-instance learning, or genetic programming or conventional statistical methods to determine whether there is the development of characteristic features of defective products. Using the trained classifier obtained with reference to any one of claims 1 to 5 to perform using a statistical method from the or from a predetermined set of features. When,
ii) signaling that no correction is required in the manufacturing process of the component when no such development is detected, or communicating a warning and / or obtaining the trained classifier Performing a corrective action in the manufacturing process of the component by modifying the process parameter via an interface with a manufacturing machine, if the process parameter was used.   7. The method of claim 6, wherein at least one defect-free component region in the examined component is extracted and the method for classifying is performed only on the defect-free component region.   The method according to any one of the preceding claims, wherein the non-destructive examination performed is a computed tomography, in particular an X-ray tomography.

Images